Automated segmentation and classification of diatoms in digital images

Diatoms are photosynthesizing algae found in almost every aquatic environment. Detection and classification of diatom species are of great importance for ecological research. Therefore, in this dissertation, the problems of automatic segmentation and classification of diatoms from digital images obtained by a light microscope have been studied. The main contributions of the dissertation are (i) a new diatom image dataset, (ii) a new segmentation model, and (iii) a new classification model. Firstly, a new annotated and publicly available diatom image dataset consisting of 2,197 color images with 3,027 diatoms from 68 species was formed to train and evaluate the models. Secondly, a novel diatom segmentation model based on edge detection and deep learning was proposed. Lastly, a lightweight but strong convolutional neural network model, namely DiatomNet, was proposed to classify diatom species. Through an extensive set of experimental studies, the performances of the proposed models were evaluated using various success metrics. The models were also compared with the recent works in the literature based on various aspects. The experimental results indicated that the proposed models surpass the recent works in almost every aspect. In conclusion, the proposed models stand out as strong candidates for automatic segmentation and classification of diatom images thanks to their promising performances.